Update chain_app.py

chain_app.py

@@ -2387,19 +2387,11 @@ async def main(message: cl.Message):
             stop=None,
         )
     
-        complete_content = ""
-    
-        # Iterate over each chunk
         for chunk in completion:
-            # Retrieve the content from the current chunk
-            content = chunk.choices[0].delta.content
-            
+            # Retrieve the content from the current chunk            
             # Check if the content is not None before concatenating it
-            if content is not None:
-                complete_content += content
-    
-        # Send the concatenated content as a message
-        await cl.Message(content=complete_content).send()
+            if chunk is not None:
+                await msg.stream_token(chunk.choices[0].delta.content)
 
     elif chat_profile == 'Llama-3-70B':
         completion = groq_client.chat.completions.create(
@@ -2417,19 +2409,11 @@ async def main(message: cl.Message):
             stop=None,
         )
 
-        complete_content = ""
-
-        # Iterate over each chunk
         for chunk in completion:
-            # Retrieve the content from the current chunk
-            content = chunk.choices[0].delta.content
-            
+            # Retrieve the content from the current chunk            
             # Check if the content is not None before concatenating it
-            if content is not None:
-                complete_content += content
-
-        # Send the concatenated content as a message
-        await cl.Message(content=complete_content).send()
+            if chunk is not None:
+                await msg.stream_token(chunk.choices[0].delta.content)
 
     elif chat_profile == 'Llama-3-8B':
         completion = groq_client.chat.completions.create(
@@ -2447,23 +2431,16 @@ async def main(message: cl.Message):
             stop=None,
         )
 
-        complete_content = ""
-
-        # Iterate over each chunk
         for chunk in completion:
-            # Retrieve the content from the current chunk
-            content = chunk.choices[0].delta.content
-            
+            # Retrieve the content from the current chunk            
             # Check if the content is not None before concatenating it
-            if content is not None:
-                complete_content += content
-
-        # Send the concatenated content as a message
-        await cl.Message(content=complete_content).send()
+            if chunk is not None:
+                await msg.stream_token(chunk.choices[0].delta.content)
 
     elif chat_profile == 'gemma2-27B':
         client = Client("gokaygokay/Gemma-2-llamacpp")
-        result = client.predict(
+        stream_list = []
+        for token in client.predict(
                 message=message.content,
                 model="gemma-2-27b-it-Q5_K_M.gguf",
                 system_message=f"You are neural nexus official chatbot, you are made by Artin Daneshvar and Sadra Noadoust and you are here to help people",
@@ -2473,10 +2450,10 @@ async def main(message: cl.Message):
                 top_k=40,
                 repeat_penalty=1.1,
                 api_name="/chat"
-        )
-        await cl.Message(
-            content=result
-        ).send()
+        ):
+            stream_list.append(token)
+        for res in stream_list:
+            await msg.stream_token(res)
 
     elif chat_profile == 'gemma2-9B':
         completion = groq_client.chat.completions.create(